The Model Was GA. My Account Wasn't.
I came to Microsoft Ignite with one learning goal: get hands-on with Microsoft Foundry.
I’ve been building AI solutions for The CTO Advisor and Layer2C for well over a year. I have a solid grasp of what it takes to build a full AI stack. Solid enough that we created the 4+1 AI Infrastructure Framework, the foundation for Layer2C assessments. Microsoft Foundry is in those assessments. But if you’ve spent any time with CTO Advisor content, you know it’s not enough for me to intellectually know what a product does. I need to feel it. I need to deploy it. I need to experience it.
So I was excited when Microsoft announced Microsoft-Decision-1.1 as generally available (GA). Microsoft built it to make structured decisions in software: routing, classification, verification, workflow control. That makes it a System One model, in the same category as TypeSafe’s Jev. A System One model doesn’t write paragraphs. Jev answers a typed question with an answer and a probability, and nothing else. That makes it a candidate for the seat where code runs the workflow and only needs a model to answer a closed question. In Lab 30 I put Jev and Articul8’s Bivia in that seat beside Gemma and Claude Haiku. Bivia alone billed about 957 million tokens, mostly because of a 1,000-token minimum per call. I wanted to experience Decision-1.1 and live in the entire Microsoft ecosystem, Copilot aside.
The goal was simple. Could I build a lab end to end on Azure infrastructure and Azure models, and get an outcome similar to my existing Clinic proof of concept? That’s the whole point of the Clinic: take a proof of concept that’s already designed, lift it, and deploy it on a new platform.
Where it broke
That’s where we hit friction. My Azure subscriptions are years old and pay-as-you-go. The monthly bill has ranged from a few dollars to about $100. Neither had used Foundry. When I registered the service, Foundry put both on its Free Tier.
On the Free Tier, most of the catalog isn’t available. In East US 2, 56 of the 92 GA models with a quota line read zero. Decision-1.1 read 0 of 0 in all 55 region and deployment-type entries, and Azure refused the deployment both times I tried. The Free Tier can’t request an increase, and there’s no API to file one. Microsoft’s quota documentation says the starting tier weighs your relationship with Microsoft. It doesn’t publish a threshold. My history wasn’t enough, and I haven’t tested what an Enterprise Agreement starts on.
So I tested all three
That made me curious. What’s the brownfield experience on each cloud for API access to foundation models? Not a fresh account built for a demo. The accounts I already use: my AWS lab account and two Google Cloud projects that carry production and test workloads.
Before the numbers: Google Cloud is a client of The CTO Advisor. Articul8 is a past sponsor; it didn’t pay for Lab 30, but at my request it raised the free-tier request limit partway through the runs. None of the vendors reviewed these labs.
The Clinic’s busiest runs peaked at about 1,300 requests and 380,000 input tokens a minute, so I set the bar at 1,300 and 400,000. A lower limit doesn’t always stop a run. It stretches it. I tested two things on each cloud. First, each cloud’s own model. Second, Claude Haiku 5.5, the Clinic’s reader and the one model all three list. If a model is going to be portable, it’s the one every cloud lists.
| Cloud | The cloud’s own model I tested | Claude Haiku 5.5 (the common model) |
|---|---|---|
| Amazon Bedrock | Nova 2 Lite answered: 2,000 requests and 8 million tokens a minute | Answered on the first call; 10 million tokens a minute |
| Google Vertex AI | Gemini 3.8 Flash accepted and billed the call: 50 million input tokens a minute, 5 billion a day | Zero quota on both projects; one increase request, on the test project, denied within 12 seconds |
| Azure AI Foundry | Microsoft-Decision-1.1 refused at deployment twice; 0 of 0 quota in all 55 region and deployment-type entries | Stopped at deployment for company details, which I didn’t submit; quota read zero |
Microsoft is the only one of the three that wouldn’t serve me the model it had just announced. And the one model every cloud lists only answered on one of them. All three list it, but only Bedrock answered.
Notice the shape of the Azure gate for Claude. Before Foundry would deploy it, it wanted my industry, organization name, and country. That’s not a technical limit. It’s a registration step the catalog doesn’t mention.
Listed isn’t the same as available. Available isn’t the same as available at your rate, either. Of the models I tested, gpt-oss-120b was the only one all three accepted. Bedrock stated 100 million tokens a minute for it. Azure stated 5 million. Vertex AI’s limit didn’t read at all. The Llama, Phi, Mistral, and DeepSeek entries on Azure sat at 20,000 tokens a minute, about 20 times under the Clinic’s bar.
It’s an account problem, not just a cloud problem
The obvious objection is that these are small accounts. An enterprise on an Enterprise Agreement might get real quota on day one. It might. I haven’t tested it, and that’s the open question.
But look at where every answer came from. Not the catalog, and not the GA announcement. The account. Quota and entitlement attach to the subscription, the project, or the account, and the cloud decides them. My AWS account has its own limits set below AWS’s defaults on other lines. Both of my Google Cloud projects read zero for Claude. And my Azure subscriptions’ history bought nothing at the Foundry door.
I didn’t test two accounts inside one organization and watch them get different answers. That’s the next lab. But the mechanism is right there.
That’s the Day 2 problem. Your proof of concept works in the account where someone already sorted out quota. Then it moves to a new subscription, a new business unit, or an acquired company’s tenant, and that account may start somewhere else entirely. How often does that happen inside your organization, and would anyone know before the architecture review?
Make one call
The full Bedrock, Vertex AI, and Foundry results are in Lab 31. Read them before you plan experiments, or plan to move labs, proofs of concept or production workloads.
The fix is cheap. Before you plan on a model through your cloud, make one real call and read its quota against your workload’s rate. Across all three clouds, that took a few hundred tokens. It won’t prove the cloud sustains your rate. It will tell you, for almost nothing, when the catalog and the GA announcement are wrong about your account.
And don’t assume portability is only a cloud-to-cloud problem. It’s also an account-to-account problem inside your own organization.
Disclosure: Google Cloud is a client of The CTO Advisor. Articul8 is a past sponsor of The CTO Advisor’s work. Neither paid for nor reviewed these labs.